Anthropic tests automated alignment agents

Anthropic said Tuesday that Claude Opus 4.6 can now act as an “automated alignment researcher,” running experiments meant to help humans supervise stronger artificial intelligence systems. (anthropic.com) The paper, published April 14, frames the problem as “weak-to-strong supervision”: a weaker teacher gives feedback to a stronger model, and researchers measure how much of the stronger model’s potential performance is recovered after training. Anthropic calls that score “performance gap recovered,” where 0 means no improvement beyond the weak teacher and 1 means the strong model reaches its best possible result. (anthropic.com) Anthropic said it used Claude Opus 4.6 to autonomously generate, test, and analyze ideas for improving that score. The company described the system as part of its Anthropic Fellows research program rather than a product release. (anthropic.com) Alignment research is the field that tries to make advanced artificial intelligence systems follow human goals and constraints. Anthropic said the issue is becoming more immediate because frontier models are already helping build their successors, including by writing large amounts of code that humans may struggle to review line by line. (anthropic.com) The new study extends a line of work Anthropic has been building for at least a year. In July 2025, the company published separate research on “alignment auditing agents,” saying automated auditors could uncover hidden goals and other concerning behaviors in large language models while scaling work that would otherwise consume large amounts of researcher time. (alignment.anthropic.com) Claude Opus 4.6 is the model Anthropic introduced on February 5, 2026, as its flagship system for coding, research, and other long-running agent tasks. Anthropic said that release added a 1 million token context window in beta and improved the model’s ability to sustain multi-step work over larger codebases. (anthropic.com) Anthropic is pitching both projects around the same operational problem: human oversight does not scale cleanly as models get more capable and are deployed in more places. Its 2025 auditing paper said alignment audits face a “scalability” problem because they require heavy expert labor, and a “validation” problem because it is hard to know what human auditors missed. (alignment.anthropic.com) Anthropic’s own framing also leaves the core tension in place: the company is testing whether artificial intelligence can help align more powerful artificial intelligence. The paper says the weak model stands in for humans and the strong model stands in for future systems that could be “much-smarter-than-human,” making the research a practical test bed for oversight before those systems arrive. (anthropic.com) For now, Anthropic is presenting the result as research on how to keep oversight from falling behind capability gains. The company’s closing question is the same one that opens the paper: whether language models can be used “to help align themselves.” (anthropic.com)